Messages in this thread | | | Subject | Re: [PATCH v2 1/4] tpm: Implement tpm_chip_find() and tpm_chip_put() for other subsystems | From | Stefan Berger <> | Date | Thu, 21 Jun 2018 16:14:46 -0400 |
| |
On 06/21/2018 03:06 PM, Jason Gunthorpe wrote: > On Thu, Jun 21, 2018 at 02:19:44PM -0400, Stefan Berger wrote: >> On 06/21/2018 01:56 PM, Jason Gunthorpe wrote: >>> On Thu, Jun 21, 2018 at 01:45:03PM -0400, Stefan Berger wrote: >>>> On 06/21/2018 01:15 PM, Jarkko Sakkinen wrote: >>>>> On Wed, Jun 20, 2018 at 04:42:33PM -0400, Stefan Berger wrote: >>>>>> Implement tpm_chip_find() for other subsystems to find a TPM chip and >>>>>> get a reference to that chip. Once done with using the chip, the reference >>>>>> is released using tpm_chip_put(). >>>>>> >>>>>> Signed-off-by: Stefan Berger <stefanb@linux.vnet.ibm.com> >>>>> You should sort this out in a way that we don't end up with duplicate >>>>> functions. >>>> Do you want me to create a function *like* tpm_chip_find_get() that takes an >>>> additional parameter whether to get the ops semaphore and have that function >>>> called by the existing tpm_chip_find_get() and the new tpm_chip_find(). The >>>> latter would then not get the ops semphore. I didn't want to do this since >>>> one time the function returns with a lock held and the other time not. >>> Another option, and I haven't looked, is to revise the callers of >>> tpm_chip_find_get to not require it to hold the ops semaphore for >>> them. >> We have tpm_chip_unregister calling tpm_del_char_device to set the ops to >> NULL once a chip is unregistered. All existing callers, if they pass in a >> tpm_chip != NULL, currently fail if the ops are NULL. (If they pass in >> tpm_chip = NULL, they shouldn't find a chip once ops are null and it has >> been removed from the IDR). I wouldn't change that since IMA will call in >> with a tpm_chip != NULL and we want to protect the ops. All existing code >> within the tpm subsystem does seem to call tpm_chip_find_get() with a NULL >> pointer, though. Also trusted keys seems to pass in a NULL pointer every >> time. >> >>> Either by giving them an API to do it, or revising the TPM entry >>> points to do it. >>> >>> I didn't look, but how did the ops semaphore get grabbed in your >>> revised patches? They do grab it, right? >> The revised patches do not touch the existing code much but will call >> tpm_chip_find_get() and get that semaphore every time before the ops are >> used. IMA is the only caller of tpm_chip_find() that now gets an additional >> reference to the tpm_chip and these APIs get called like this from IMA: >> >> ima init: chip = tpm_chip_find() >> >> ima::tpm: tpm_chip_find_get(chip) ... tpm_put_ops(chip) >> >> ima::tpm: tpm_chip_find_get(chip) ... tpm_put_ops(chip) >> >> [repeat] >> >> ima shutdown: tpm_chip_put(chip) > Maybe just change tpm_chip_find_get() into tpm_get_ops(chip) and > convert all callers?
And then re-introduce tpm_chip_find_get() for IMA to call ?
| |